You can speak English to your phone, and it writes your words. If the words are wrong, maybe your pronunciation needs work.

Researchers looked at 15 studies about this kind of practice. It helped learners, but only by a medium amount. It worked well for single sounds, like L and R. It worked better when students practised with friends than alone. So use the phone, but also ask a friend to listen.

Automatic speech recognition, or ASR, is the technology that turns your voice into text. Many teachers are now interested in using it for pronunciation practice, because learners can check whether their words are recognized. A 2024 meta-analysis of 15 studies found that ASR practice had a medium effect on pronunciation. It was most useful for individual vowels and consonants, but less useful for rhythm and intonation. Short programs did no better than practice without ASR, and learners who practiced with classmates improved much more than those who practiced alone.

However, ASR does not always hear speech the way people do. In one study, the machine's results matched human listeners' judgments for only some speakers. Learners should therefore use ASR as a practice partner and check important points with a human listener.

Speaking to a phone and checking whether it writes the intended words has become a popular way to practise pronunciation. Researchers call the underlying technology automatic speech recognition (ASR), and a 2024 meta-analysis of 15 studies, covering 38 effect sizes, estimated its overall effect on ESL and EFL pronunciation as medium. The details matter, however.

ASR with explicit corrective feedback was more effective than simple dictation, and gains were large for segmental features such as vowels and consonants but small for suprasegmental features such as stress and intonation. Short interventions did not outperform non-ASR practice, and using ASR with peers produced a large effect while using it alone produced only a small one.

A study of 84 Chinese university students supports this: all groups improved, but groups involving peer feedback outperformed the group using ASR alone. A further caveat concerns accuracy. When Google's ASR and twelve native listeners transcribed the same learners, their overall scores were similar, yet their judgments correlated for only some speakers. Machine recognition, in other words, is not a perfect stand-in for human intelligibility. The best approach seems to combine ASR with human listeners.